# Citation Needed: Provenance for LLM-Built Knowledge Graphs — Daniel Chalef, Zep AI

Source: [AI Engineer](https://www.youtube.com/watch?v=H7puB0RwJMM)  
Feed7 permalink: https://feed7.dev/p/citation-needed-provenance-for-llm-built-knowledge-graphs-daniel-chalef-1iob5t8  
Published: 2026-07-23T04:00:02.000Z  
Trust: Source Linked (source_linked)

## Why Included

LLM synthesis breaks simple source pointers when facts merge, change, or outlive one input. Model provenance inside the context graph so agents can filter trust, debug derivation, and apply deletion policies.

## Source Summary

Graphiti stores source episodes and links derived entities and facts back to them. When entities merge or new evidence contradicts a fact, those links and temporal changes preserve **fact-level lineage** instead of relying on a later log.

## Practical Implication

Build provenance into an agent-memory data structure at ingestion. Propagate source metadata, filter retrieval by trust policy, and use remaining source links to decide whether a derived fact survives **source deletion**.

## Agent-Ready Context

Graphiti stores source episodes and links derived entities and facts back to them. When entities merge or new evidence contradicts a fact, those links and temporal changes preserve **fact-level lineage** instead of relying on a later log.

Build provenance into an agent-memory data structure at ingestion. Propagate source metadata, filter retrieval by trust policy, and use remaining source links to decide whether a derived fact survives **source deletion**.

Verification policy is contextual: some decisions require any verified parent, while others require every parent to be verified. Graph construction is also **expensive in cost and latency**, and one discussed many-edge case uses a separate structure rather than keeping all provenance in the graph.

## Context Map

- Layer: context
- Domains: data, security
- Topics: agent-memory, retrieval, context-engineering

## Uncertainty

- Verification policy is contextual: some decisions require any verified parent, while others require every parent to be verified. Graph construction is also **expensive in cost and latency**, and one discussed many-edge case uses a separate structure rather than keeping all provenance in the graph.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.